Re: [PATCH 1/3] sched/headers: Fix compilation error with GCC 12

[Christophe de Dinechin]

> On 14 Apr 2022, at 17:21, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> On Thu, Apr 14, 2022 at 05:08:53PM +0200, Christophe de Dinechin wrote:
>> With gcc version 12.0.1 20220401 (Red Hat 12.0.1-0) (GCC), the following
>> errors are reported in sched.h when building after `make defconfig`:
> 
> <snip tons of noise>

I don’t mind removing the detailed error message.
What do others think?

> 
>> Rewrite the definitions of sched_class_highest and for_class_range to
>> avoid this error as follows:
>> 
>> 1/ The sched_class_highest is rewritten to be relative to
>>  __begin_sched_classes, so that GCC sees it as being part of the
>>  __begin_sched_classes array and not a distinct __end_sched_classes
>>  array.
>> 
>> 2/ The for_class_range macro is modified to replace the comparison with
>>  an out-of-bound pointer __begin_sched_classes - 1 with an equivalent,
>>  but in-bounds comparison.
>> 
>> In that specific case, I believe that the GCC analysis is correct and
>> potentially valuable for other arrays, so it makes sense to keep it
>> enabled.
>> 
>> Signed-off-by: Christophe de Dinechin <christophe@xxxxxxxxxxxx>
>> Signed-off-by: Christophe de Dinechin <dinechin@xxxxxxxxxx>
>> ---
>> kernel/sched/sched.h | 11 +++++++++--
>> 1 file changed, 9 insertions(+), 2 deletions(-)
>> 
>> diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
>> index 8dccb34eb190..6350fbc7418d 100644
>> --- a/kernel/sched/sched.h
>> +++ b/kernel/sched/sched.h
>> @@ -2193,11 +2193,18 @@ const struct sched_class name##_sched_class \
>> extern struct sched_class __begin_sched_classes[];
>> extern struct sched_class __end_sched_classes[];
>> 
>> -#define sched_class_highest (__end_sched_classes - 1)
>> +/*
>> + * sched_class_highests is really __end_sched_classes - 1, but written in a way
>> + * that makes it clear that it is within __begin_sched_classes[] and not outside
>> + * of __end_sched_classes[].
>> + */
>> +#define sched_class_highest (__begin_sched_classes + \
>> +			     (__end_sched_classes - __begin_sched_classes - 1))
>> #define sched_class_lowest  (__begin_sched_classes - 1)
>> 
>> +/* The + 1 below places the pointers within the range of their array */
>> #define for_class_range(class, _from, _to) \
>> -	for (class = (_from); class != (_to); class--)
>> +	for (class = (_from); class + 1 != (_to) + 1; class--)
> 
> Urgh, so now we get less readable code,

You consider the original code readable? It actually relies on a
precise layout that is not enforced in this code, not even documented,
but actually enforced by the linker script.

> just because GCC is being
> stupid?

I think that GCC is actually remarkably smart there. It tells you
that you are building pointers to A[] from B[], when there is a legit
way to say that the pointer is in A[] (which is what my patch does)

> What's wrong with negative array indexes? memory is memory, stuff works.

What’s wrong is that the compiler cannot prove theorems anymore.
These theorems are used to optimise code. When you write -1[B], the
compiler cannot optimise based on knowing this refers to A[B-A-1].

While at first, you might think that disabling a warning is a win, what comes next
is the compiler optimizing in a way you did not anticipate, mysterious bugs showing up,
and/or having to turn off some potentially useful optimisation.